Modeli digitalne predistorzije za hibridne masivne višeantenske predajnike sa formiranjem snopa primenom neuralnih mreža

Abstract

Intenzivan i brz razvoj bežičnih sistema nove generacije usko je povezan sa razvojem i primenom Multiple-Input Multiple-Output (MIMO) tehnika koje omogućavaju povećanje protoka i spektralne efikasnosti, kao i pouzdanosti sistema. Sa druge strane, javljaju se novi izazovi pri dizajniranju masivnih višeantenskih (massive MIMO – mMIMO) predajnika zbog pojave nelinearne distorzije signala. Kako bi se smanjila nelinearna distorzija signala i postigle što bolje performanse MIMO sistema, neophodno je posebnu pažnju posvetiti digitalnoj predistorziji (DPD) pojačavača kod mMIMO predajnika. U okviru ove doktorske disertacije realizovani su i analizirani različiti DPD modeli za mMIMO predajnike sa hibridnim formiranjem snopa (hybrid beamforming) primenom neuralnih mreža, sa ciljem razvijanja što efikasnijeg i sofisticiranijeg DPD modela. Neuralne mreže (NN) su prvenstveno izabrane zbog svoje sposobnosti da veoma dobro aproksimiraju nelinearne funkcije, kao i zbog svoje prilagodljivosti. Predložen je efikasan Real‐Valued Time‐Delay Neural Network with 2 hidden Layers (RVTDNN2L) DPD model, kao i proširen RVTDNN2L DPD model kod koga se u cilju povećanja tačnosti modela koristi dodatni signal koji sadrži informacije o koeficijentima beamforming-a. Predloženi modeli su implementirani u programskom paketu Matlab i nakon sprovedenih sveobuhvatnih simulacija, izvršena je analiza i verifikacija efikasnosti kompenzacije nelinearne distorzije. Simulacije su izvršene za 64x64 HBF mMIMO sistem sa jednim korisnikom i sa više korisnika, pri čemu su korišćene izmerene vrednosti pojačavača. Dobijeni rezultati prikazani su korišćenjem grafika i numerički korišćenjem relevantnih metrika: normalizovane srednje kvadratne greške (NMSE) i amplitude vektora greške (EVM). Kao rezultat sprovedenog istraživačkog rada, pokazano je da predloženi RVTDNN2L DPD model i prošireni RVTDNN2L DPD model znatno bolje kompenzuju nelinearnu distorziju u odnosu na polinomijalne modele, kao i u odnosu na ostale razmatrane NN DPD modele slične kompleksnosti, čime je izvršeno fundamentalno unapređenje efikasnosti kompenzacije nelinearne distorzije signala u bežičnim sistemima nove generacije.

The intensive and rapid development of new generation wireless systems is closely related to the development and application of Multiple-Input Multiple-Output (MIMO) techniques, which increase the throughput and the spectral efficiency, as well as the reliability of a system. On the other hand, there are new challenges when designing massive MIMO (mMIMO) transmitters due to nonlinear signal distortion. In order to reduce the nonlinear signal distortion and achieve the best possible performance of a MIMO system, it is necessary to pay special attention to the digital predistortion (DPD) of amplifiers in mMIMO transmitters. Within this doctoral dissertation, different DPD models for mMIMO transmitters with hybrid beamforming using neural networks are implemented and analyzed, with the aim of developing the most efficient and sophisticated DPD model. Neural networks (NN) were primarily chosen because of their ability to approximate nonlinear functions very well, as well as because of their adaptability. An efficient Real-Valued Time-Delay Neural Network with 2 hidden Layers (RVTDNN2L) DPD model is proposed, as well as an extended RVTDNN2L DPD model where, in order to increase the accuracy of the model, an additional signal containing information about the beamforming coefficients is used. The proposed models are implemented in the Matlab software package and after comprehensive simulations, the analysis and verification of the effectiveness of the nonlinear distortion compensation is performed. Simulations were performed for a single-user and multi-user 64x64 HBF mMIMO system, based on measurement data from an actual amplifier. The obtained results are presented graphically and numerically using relevant metrics: normalized mean square error (NMSE) and error vector amplitude (EVM). As a result of the research work, it is shown that the proposed RVTDNN2L DPD model and the extended RVTDNN2L DPD model compensate for nonlinear distortion significantly better in comparison to polynomial models, as well as in relation to other considered NN DPD models of similar complexity, by which fundamental improvement in the efficiency of compensation of nonlinear distortion signals in new generation wireless systems has been achieved.